Fix benchmark_inference.py segfault and superficial errors with multiple devices

[shino16]

• Makes _call_dispatch use __get__(slf, type(slf)) instead of __get__(slf), working around a CPython segfault (python/cpython#132747). From Python 3.13.3 this workaround will no longer be necessary, but I think we can keep it anyway.
• Creates TensorProxy.shape's provenance only when the tensor's history exists, so that incomplete provenance record will not be created. This fixes the AssertionError in mentioned in #2677 (comment).
• Makes thunder/core/update_aliases.py skip numel checks (introduced in #2716) when relevant shapes are dynamic.

Multi-device support of benchmark_inference.py is still incomplete. Refer to #2677 for progress.

[beverlylytle]

Solving #2677 completely is going to blow this PR up into a monster, touching lots of files. I think it would be better to fix each issue individually. It will make it easier to choose targeted reviewers for each change. This will also potentially save time later, if an issue pops up and bisection is used to find the culprit commit. So, let's just start with the seg fault fix.

[shino16]

Your suggestion makes sense, I'll submit separate small PRs. And thank you for your reviews in #2746! @beverlylytle